The web has a new reader

The new reader

Speakers on both days of sessions said buying now starts inside an AI assistant, where a buyer learns, compares and chooses in one conversation. They showed fewer clicks from search, sessions that never reach the open web, bots that account for half of page views, and an agent that paid for a purchase.

The sessions. Eric Stine, Sitecore, mainstage keynote. Michelle Boockoff-Bajdek, Sitecore, mainstage keynote. Both on Wednesday, November 11. Chris Andrew, Scrunch, mainstage. Nilesh Tamhane and a colleague, XCentium, breakout. Kiran Patil and Khushboo Sorthiya, Horizontal Digital, breakout. Kevin Schofield, Velir x Brooklyn Data, panel. Andrew Michael, Scrunch, breakout. All on Thursday, November 12.

The journey starts in a conversation

Eric Stine, CEO of Sitecore, the company whose conference this was, told a story of his own in the opening keynote. In August 2026 he volunteered to coach his son's flag football team and asked ChatGPT how to run a practice. The conversation led him to playbooks, wristbands, cones and flags from brands he had no idea were in that business. He said he spent more than $1,500 without ever setting out to buy anything.

His lesson was that awareness, consideration and intent have been "scrunched together" into one conversation, and that a brand which had the right answer but never appeared in it was never considered at all.

Michelle Boockoff-Bajdek, Sitecore's Chief Marketing Officer, described shopping for a mortgage in a conversation with ChatGPT. For the bank's marketer, she said, none of it was visible: no ad click, no page view, no form.

A funnel in one window

Andrew Michael, VP of product at Scrunch, a company that tracks how brands appear in AI answers, pictured marketing as a set of stages. His claim is that they are collapsing into a single interface, where what took days, weeks or months now happens within a minute: the assistant reads the sources, compares the products and in some cases makes the purchase.

Fewer clicks from search

Kiran Patil and Khushboo Sorthiya of Horizontal Digital, a digital agency, showed Pew Research Center's count of what an AI summary does to clicks on Google. The count rests on 900 U.S. adults and 68,879 of their searches: on how many visits to Google did the person click a link? The slide's line is that buyers have not stopped researching, they have stopped clicking.

Nilesh Tamhane and a colleague from XCentium, a Sitecore partner, showed a slide on the same subject headed "Search is changing". One of its cards says 60% of searches end without a click. Another is headed "58% drop in organic traffic", which means visits that arrive from unpaid search results. Its small print credits the 58% to Ahrefs as a drop in CTR once an AI Overview, Google's AI summary, appears. CTR is click-through rate: the share of people shown a link who click it.

Sessions that never leave the chat

Boockoff-Bajdek said on Wednesday that more than 41% of AI sessions happen entirely inside the AI platform, with no open-web search before and no click to the open web after. On Thursday Chris Andrew, Scrunch's co-founder and CEO, showed where the rest go. His slide drew four paths across three columns: the web, the AI assistant, and the web again afterwards. His conclusion was that these journeys are condensing, and that the website has never been more important.

A third audience

Kiran Patil said sites were built for two audiences, people and search engines, and that a third has arrived: the programs that fetch pages for AI systems. Kevin Schofield of Velir x Brooklyn Data cited Cloudflare as saying that in June 51% of all page views and hits came from bots and not from people.

Agents that finish the job

Chris Andrew described handing purchases to Meta's Muse agent. Asked to find an out-of-print book, it searched the web, found a $12 copy, asked only for a shipping address and paid with a card Meta already held.

Tickets, and a $500 deposit on a car the agent had recommended, followed. His sentence for it: he expressed intent and the agent delivered an outcome. Of the tickets he said he did not visit a website.

Found is not chosen

Tim Sanders, Chief Innovation Officer at G2, opened the second day with a research talk on how AI assistants decide which products to recommend. He said buyers now reach a shortlist of three inside a single prompt, and he defined winning as being one of the three.

The sessions. Tim Sanders, G2, mainstage keynote and the marketing leaders panel. Nilesh Tamhane and a colleague, XCentium, breakout. Andrew Michael, Scrunch, breakout. All on Thursday, November 12.

Buyers now start in chat

G2, a software review site, asked B2B software buyers where they start when they research a purchase. In April 2025, 29% said they start with AI search more often than with Google. In March 2026 the answer was 51%. He said the number had jumped 76% in a year (51 is about three quarters more than 29).

Sanders said the question set a Google property against ChatGPT, Gemini, Perplexity, Claude or DeepSeek, so the AI Overviews above Google's results do not count toward the 51%. He said G2 ran the March survey twice because he did not believe the first result. On stage he added two shares that are not on the slide: 58% of large enterprise buyers and 65% of IT professionals start in chat.

The shortlist is made in one prompt

The more important change, he said, is that buyers have handed the gatekeeper's job to the assistant. G2 has studied for about a decade how a buying committee reaches its shortlist of three, and he put that work at 6 to 15 working hours: search results, analyst reports, peers, reviews, a spreadsheet. Buyers now get there in one prompt, usually without realizing it. They describe the job to be done and their constraints, and the model returns three options with caveats.

He added a figure he attributed to 6sense, a company that does similar research: 93% of buyers decide from the shortlist they had on day one. The job, he said, is to win the answer, which he defined as being one of the three recommended. Showing up or being cited is not the same thing.

Three problems

Sanders built the rest of the talk on three problems, each with a question on his slide and a time horizon he gave on stage.

The prompt problem

He said AI visibility reporting runs on a prompt panel, the set of prompts a brand tracks, and that panels are usually built by turning keywords, category names and the brand's own product language into questions, so they are guesses, and model vendors do not release real prompts. The worst effect he named is a false positive: a brand that tracks prompts in its own vocabulary sees a very high citation share of voice, meaning it is cited in a large share of the answers to prompts it wrote itself.

His example was two competitors in one software category. One was promoting a new category name, AI visual collaboration software, while its traffic and pipeline fell. The other was optimizing for the word "whiteboards", because its business development reps had been collecting real prompts from prospects and that was the word buyers used.

What a real prompt asks for

Agents, Sanders said, are loyal to the prompt: they break it into components and treat each as a box to check before a brand is a candidate. His slide set a Google search beside what he said was a real captured prompt, with three phrases of the prompt highlighted.

He said each highlighted requirement has to be answered by the vendor's own product documentation, or the vendor is not on the long list. A vendor whose site says "hospitals" 132 times but documents neither reimbursement speed nor mobile use may be left out, he said. Later he pictured the long list for this prompt as 14 vendors that had checked every box.

In thousands of prompts captured from his customers' buyers, the job and its constraints made up almost 60% of the words in a prompt, and the category name (here, "CRM") appeared in only two prompts of every three.

He made one claim about pricing. Citing a study he attributed to Boston University and Northwestern, he said models are tuned to recognize patterns of manipulation, and that hiding basic terms such as price or contract length behind a talk-to-sales form signals that risk and can get a vendor left out.

His fix is to weight real buyer language above invented prompts. The gold standard he named is prompt capture: asking a prospect who arrived through an assistant for the prompt they used.

Mention, citation, recommendation

Sanders opened the second problem with one line: a company whose pipeline is suffering does not have an AI visibility problem, it has an evidence problem. He then ranked the ways a brand can appear in an answer, and gave three of them a click-through rate: of the people who are shown the answer, the share who click through to the brand's site.

His reason for the low rate on a citation was that the whole point of AI search, for the person using it, is not having to click. He compared counting citations to counting website hits in 2001, citing a September study by Jeff Oxford: over thousands of runs of commercial-intent prompts, a brand that was cited was still left off the list of recommendations 60% of the time.

XCentium's presenters separated the first two outcomes that afternoon. Of 100 people asking a question the brand wants to answer, Nilesh Tamhane said, 16 hear its name and 6 get a link to its site. The other 94 get no path to the site.

Source. Nilesh Tamhane, XCentium, on stage. XCentium's slide reads "16% mention your brand" and "6% of those link to you", and credits a Webflow analysis of 2,000 U.S. company sites.

One question, two kinds of answer

Andrew Michael, VP of product at Scrunch, a company that tracks how brands appear in AI answers, said Sanders had been right about this distinction and that Scrunch sees it in its data. His slide shows one question answered two ways.

Evidence that matches the claim

Sanders gave a three-step model of how an assistant reaches a recommendation. He said it comes from academic studies that record a model's reasoning across many runs.

He said the evidence has to match the claim. A security company he worked with knew from its sales data that it won 80% of its deals when the prospect believed its agent latency was 25 milliseconds. Its analyst report, reviews and testimonials praised ease of use, quality and service, and none mentioned latency. He added that G2's research found a third of buyers had bought from a vendor they had never heard of before an assistant named it.

One of his last slides is titled "ChatGPT's Recommendation Algorithm" and divides what drives a ChatGPT recommendation among five factors. He called it the best thing he has seen for anticipating the algorithm.

His reading was that a place on a high-authority list matters most for what he called lazy prompts and can break a tie, while awards, reviews and customer examples are gaining influence because they are specific.

Where the budget goes

Sanders said that for every dollar most B2B companies spend on answer engine optimization (AEO) in 2026, 40 go to SEO tools, services, content and search ads. "1 to 40 means 2.5%," he said: AEO spending is one fortieth of SEO spending. He did not argue for an even split, only that 2.5% is wrong. His slide splits an AEO budget five ways.

On the panel later that morning he said website traffic is no longer a meaningful measure, and offered two others: have reps ask every inbound lead how they heard of the company, and track year-over-year visits to the contact page.

How an assistant picks its sources

Five sessions explained some part of how an AI assistant decides what to read and cite. Together they follow one sequence, from a buyer's question to the handful of pages an assistant quotes, and they describe a reader that does not scroll, click or run scripts.

The sessions. Kiran Patil and Khushboo Sorthiya, Horizontal Digital, breakout. Devin Stevens, Scrunch, breakout. Nilesh Tamhane and a colleague, XCentium, breakout. Andrew Michael, Scrunch, breakout. Chris Andrew, Scrunch, mainstage. All on Thursday, November 12.

One question becomes several searches

Kiran Patil of Horizontal Digital, a digital agency, showed the sequence as five steps on a slide. A buyer asks a question. The assistant does not search for the question as typed: it breaks it into several search queries. A search index answers with about twenty candidate pages. The assistant reads only three to ten of them, and those are the pages that shape the answer and receive the citation, the link an assistant gives as the source of its answer.

Andrew Michael, VP of product at Scrunch, a company that tracks how brands appear in AI answers, put a number on the second step. Behind one question, he said on stage, the model runs another twelve searches that the person asking never sees.

No index entry, no candidacy

The index in the third step is an ordinary search index. Horizontal Digital's slide names Google, Bing and Brave, and its conclusion is the heading above: a page that is not in the index cannot be a candidate, and a page the crawler cannot read cannot be one of the three to ten. A crawler is a program that fetches pages for a search engine or an AI system. Patil applied this to the three acronyms now in use, which his next slide set in a row: search engine optimization (SEO), answer engine optimization (AEO) and generative engine optimization (GEO).

The newest visitor

Horizontal Digital opened its session with one line from a web server's log: a request from GPTBot, OpenAI's crawler, with a caption on what this visitor does not do.

x.x.x.x - - [12/Nov/2026:14:40:02] "GET /services/wealth HTTP/1.1" 200 "Mozilla/5.0 AppleWebKit/537.36 (KHTML, like Gecko; compatible; GPTBot/1.1; +https://openai.com/gptbot)"

Three kinds of bot

Horizontal Digital publishes an open-source app, "Is It Agent-Ready?", and the app's documentation sorts bots into three kinds. It leaves the blocking of a training bot to the site's owner.

What the bot receives

A later slide of Horizontal Digital's, headed "What the bot actually sees", puts two panels side by side: a page as its visitor sees it, and what GPTBot receives when that page is built in the browser, which is an empty div and a script tag. The slide states it as a rule: most AI crawlers fetch the HTML and do not run the JavaScript.

<html>
  <head><title>BrandX</title></head>
  <body>
    <div id="root"></div>
    <script src="/static/main.js">
    </script>
  </body>
</html>

Patil's test for any page needs no tool. Open it in a browser, choose view source and look for the content. If it is not in the source, a bot will not see it.

Nilesh Tamhane and a colleague from XCentium, a Sitecore partner, gave the everyday cases of content a bot can miss: an FAQ inside an accordion that is closed by default, tabs, carousels, drop-downs and video.

A limited appetite

Two speakers said a crawler spends only a fixed amount on a page. Tamhane put it at about 10,000 tokens a page, after which, he said, a bot moves on. Patil said a crawler processes only a set number of tokens. A token is the unit a language model counts text in: a short word, or a piece of a longer one.

Source. About 10,000 tokens: Nilesh Tamhane, XCentium, answering an audience question. Asked where the figure comes from, he said he thought it is on Scrunch's website.

The same question, different sources

Devin Stevens of Scrunch opened with two screen recordings of one question, put to ChatGPT and to Perplexity minutes apart. His point was that these systems are probabilistic, so one answer is not a measurement.

He then showed a month of such questions, about a thousand answers from each assistant, with the sites those answers cited sorted into types. The questions were non-branded: they named no brand.

The slide's own summary puts McDonald's and the other restaurant brands together: 58% of ChatGPT's citations went to restaurant brands' own sites, against 18% of Perplexity's. Stevens's conclusion was that the split of effort between a brand's own pages and other people's depends on which assistant its customers use.

Most citations are someone else's page

Chris Andrew, Scrunch's co-founder and CEO, said on the mainstage that 70 to 80% of the content an assistant uses to check an answer comes from outside the brand's own site. Andrew Michael said more than 80% of citations are to third-party sources, which are sites the brand does not own.

Before any data: four decisions

Devin Stevens of Scrunch, a company that tracks how brands appear in AI answers, named four things a team settles before a tracking tool runs its first prompt: the brand, the competitors, the personas and the prompts. Speakers from G2 and XCentium said where their own prompts come from.

The sessions. Devin Stevens, Scrunch, breakout. Tim Sanders, G2, mainstage keynote. Nilesh Tamhane and a colleague, XCentium, breakout. All on Thursday, November 12. Lisa Saltsman, Icreon, Jaina Baumgartner, Pixleration, and Rodrigo Peplau, MSQ DX, panel, Wednesday, November 11.

What is being set up

Stevens's slide gives a monitoring program five parts. Seven to ten prompts a topic is usual, he said. The slide's other line is that the data that comes out is only as good as the prompts that go in.

His session was planned around five decisions. A fire alarm ended it about 28 minutes into its 45, after the fourth.

The brand

The brand came first. Stevens's example throughout was McDonald's, which Scrunch had tracked as a demonstration. Its names include "Mickey D's" and "The Golden Arches". To the corporate, newsroom, careers and regional sites on his slide he added support pages, FAQs, pages behind a sign-in and the company's LinkedIn profile. An assistant asked whether a company is growing, he said, will read its careers page.

The competitors

Many brands, Stevens said, arrive with three competitors from a sales comparison deck, usually the ones they want to compete with. The list to use, his slide says, is the brands named in answers to non-branded questions, which are questions that do not name the brand. The slide counts about 8,400 AI responses to such questions in his demonstration, on all the platforms tracked, from September 27 to October 26, 2026.

For a quick family meal, he added, the competition includes a supermarket's rotisserie chicken.

The personas

An assistant answers using what it knows about the person asking, so the method adds a persona's context to each prompt. Stevens showed three.

Where good prompts come from

Stevens's slide names four sources of prompts and gives an example of each. The fourth is the brand's own website, which the slide says shows the territory and is not to be mirrored by the prompts.

Tim Sanders of G2, a software review site, put real buyer language first, as the section on his keynote reports: reviews, analyst reports, recorded sales calls and requests for proposal. He said chief marketing officers have told him 100 real prompts change everything.

XCentium, a Sitecore partner, builds its prompts around four stages of a purchase. The playbook handed out at Wednesday's partner panel splits prompts by persona and tags them by stage of the sales funnel.

Good prompt, bad prompt

Another of Stevens's slides names three mistakes: a prompt too short to carry intent, a prompt that mirrors the brand's own website, and several rewordings of one question. Of the second he said a set built from questions such as why the brand is the best solution will show the brand every time and tell its team nothing.

A healthy prompt set

In the healthy set on Stevens's slide, about 80% of the prompts are non-branded and measure competition for demand, and about 20% name the brand and measure reputation and accuracy. The set spans five kinds of question and every stage of the customer's journey.

His slides then audit the demonstration's own 990 prompts, under the caption "Heavy on discovery. Nothing at the moment of choice."

From finding to fix

Speakers from XCentium, Velir x Brooklyn Data and Pixleration described what is done between reading an AI visibility report and seeing a page cited: how a gap is diagnosed, what is changed, and how the result is checked. XCentium showed a client's numbers before the work and four weeks into it.

The sessions. Nilesh Tamhane and a colleague, XCentium, breakout. Kevin Schofield and David Gelman, Velir x Brooklyn Data, panel. Both on Thursday, November 12. Lisa Saltsman, Icreon, Jaina Baumgartner, Pixleration, and Rodrigo Peplau, MSQ DX, panel, Wednesday, November 11.

A cycle, not a one-time project

Nilesh Tamhane and a colleague from XCentium, a Sitecore partner, presented the way they run AI visibility work for a client as four steps. The heading of their slide is that AI visibility is an optimization cycle and not a one-time project.

The cycle starts from the gap Tamhane described, reported in the section on Tim Sanders's keynote: of 100 people who ask AI a question the brand wants to answer, 16 hear its name and 6 get a link to its site. His colleague said the aim is not more citations for their own sake, but more chances to influence the buyer's decision.

The baseline: where one client stood

The first step begins with the decisions reported in the section on what is settled before any data, and ends with a baseline. XCentium's came from a real client, shown as a tracking report for the week of August 24 to 30, 2026, before any change was made: 200 AI responses, with five competitors and four assistants tracked.

The presenter read the first two numbers together: the client is mentioned in about 70% of the tracked conversation and cited in about 20%, and the difference is visitors who had no link to follow.

Five causes behind one symptom

XCentium's slide for the analyze step gives one symptom five possible causes.

A book in a cupboard with no label

Tamhane said the cause of the gap is usually not the content itself. His image was the best book in the library, put in a cupboard with no label. His slide on the gap gives four reasons. Structured data is the machine-readable facts about a page, and llms.txt is a file written for agents.

In the implementations XCentium has done over the last four months, he said, the issue has mostly been accessibility, JavaScript, how content is served over the CDN (the network that delivers a site's files) or how the site's components are set up. Kevin Schofield of Velir x Brooklyn Data said the same of his own work that afternoon: content is usually not the problem and access is.

A fix for each finding

The slide for the act step pairs three findings with a diagnosis and an action.

As spoken, the fixes were answer-first content, a description that answers a question and is separate from the standard meta description, clearer headings, more concise summaries, and a structure made for how a bot reads a page. Both presenters said the answer is usually to improve existing content, not to publish a new page each time a prompt goes uncited.

Hidden content, and a second way out

Asked why tabs and accordions are a problem, XCentium's presenters said some bots cannot open content that sits behind JavaScript, such as an FAQ accordion collapsed by default. They recommended putting the same content in structured data (as JSON-LD, a block of data in the page's HTML), which a bot can read without interacting with the page. They said the same of carousels, drop-downs and video, where a transcript belongs on the page and in the JSON.

At the Velir panel a questioner said accordions hide content from agents and help page design. Schofield answered that a website is a restaurant where people are served, and an assistant only wants delivery. The content needs a second way out, he said, such as a Markdown version, and people's experience should not be made worse for bots.

Jaina Baumgartner of Pixleration drew a line on Wednesday's panel. Creating different content for AI engines is cloaking, she said, and it gets a site left out of results. Agent experience pages, her term, are not that: they serve the engine the same content as the person, in a form it need not work to read.

The same work, as a handout

A two-page playbook from Scrunch was handed out at Wednesday's partner panel of Icreon, Pixleration and MSQ DX. It lists similar work: lead with the direct answer, then clear headings, definitions, comparison tables and short summaries. It adds refreshing old pages that are still cited, rewriting pages that agents visit but do not cite, and checking that AI crawlers are not blocked. It calls fixing factual errors about the brand an easy early win, and says that for authority third parties usually carry the most weight.

Four weeks on the same prompts

XCentium works on one subject area, product group or business area at a time. Its presenters showed the client of the baseline four weeks on, with the brand names removed. The measure is again citation share: the percentage of AI responses that cite the brand's content, set beside the same percentage for its competitors' content.

The presenters asked themselves how they know the changes moved the number. Tamhane's answer was the method: take 50 or 100 prompts, set a baseline, make changes aimed at those prompts and keep tracking the same ones.

Asked how long it takes, they said about four weeks after the changes shows where things stand, and another four to six weeks may be needed.

Choosing the measure first

David Gelman of Velir x Brooklyn Data said a visibility tool gives a view without context, and that the measure has to be chosen before the work starts. This section reports what he and other speakers said a visibility number needs beside it, and who owns the work.

The sessions. Kevin Schofield and David Gelman, Velir x Brooklyn Data, a panel without slides. Kiran Patil, Horizontal Digital, breakout. Andrew Michael, Scrunch, breakout. All three on Thursday, November 12. Lisa Saltsman, Icreon, and Jaina Baumgartner, Pixleration, partner panel, Wednesday, November 11.

Numbers without context

Gelman, who described himself as a data practitioner, said marketing effectiveness has never been judged by one method. He named three that inform each other and called them a tripod. AI visibility, he said, makes it a chair with four legs.

What Scrunch, a company that tracks how brands appear in AI answers, or any tool of its kind gives, he said, is a good view and a limited one, because it operates without context. The context is data the business already holds: margin, the CRM (customer and sales records), the ERP (orders and finance), personas and segmentation. Joined to those, he said, the numbers say what the visibility is worth to the business.

What a referral report counts

Kiran Patil of Horizontal Digital, a digital agency, showed a referral report for his firm's own site, covering mid August to November 12, 2026. It counted 535 sessions that arrived from an AI platform. Of those, it counted 322 as engaged, meaning the visitor stayed and interacted, and 22 as ending in a conversion. A conversion in this report is one of three events: a video event, a form view or a form submission.

Chosen before the work starts

The panel's moderator asked how to tell what is working when AI does not send clean referral traffic. Gelman said the measure has to be chosen before the work starts: revenue, marketing qualified leads or share of voice. With it go a hypothesis, a time limit and the size of change that would count as an effect. Without other data beside the visibility numbers, he added, a team cannot show even a correlation.

How long the joining takes

Whether connecting that data takes years depends on the starting point, Gelman said. With a mature data practice it can be done in weeks, because the joins are on topics, pages and products, and no one has to match people or companies across systems. Even an extract of product margins is often enough to start.

Kevin Schofield, also of Velir x Brooklyn Data, added that SEO takes one to six months before a page ranks, and that citations in an assistant can change in as little as four days.

Keeping the data

Gelman said Scrunch currently keeps 90 days of history, and advised pulling the data out to set a baseline: what is not captured cannot be recovered later.

That afternoon an audience member asked Andrew Michael, Scrunch's VP of product, about a 90-day rolling retention. Michael said the data can be exported and backed up, that he was not sure about 90 days and was pretty sure retention is much longer, and that the team would follow up with exact periods.

Source. 90 days of history: David Gelman, on stage. The answer on export and retention: Andrew Michael, Scrunch, to an audience question in his own session.

Who owns the work

Gelman also said someone among IT, data, marketing and the web team has to drive the process and own the analysis, with the authority to act and not only the responsibility.

On Wednesday's partner panel, Lisa Saltsman of Icreon polled the room on who owns AI visibility, marketing or IT, and most hands went up for a combination. Jaina Baumgartner of Pixleration said the most successful organizations she works with have a digital team between the two.

The agent-ready checklist

Horizontal Digital, a digital agency, presented an open-source app that scores a web page on whether an AI agent can reach, read, understand and trust it. This section lists its 20 checks on a page and its 12 checks on a site, each with what passes it and its tier, and says how a page is scored.

The session. Kiran Patil, Group Director, Technology, and Khushboo Sorthiya, Technology Lead, Horizontal Digital, breakout. Thursday, November 12.

The four questions

The app is called "Is It Agent-Ready? for SitecoreAI". It is open source and installs in SitecoreAI, the Sitecore product in which editors build pages, where it scores a draft page before the page is published. The slide that introduced it puts four questions at its center, and every page check sits under one of them.

The same slide names what the app draws on: the page's fields and layout, the page's HTML fetched the way each AI bot fetches it, without JavaScript, and an LLM (a large language model, the kind of model behind an AI assistant) that drafts fixes such as a title and description, FAQ markup, an image's written description and an author.

The app's public documentation lists every check with its pass mark. Every check, tier, pass mark and scoring rule in this section is from that documentation as it stood on November 12, 2026: the README and the technical list of page checks in the app's repository. The pass marks are restated here, and the thresholds are Horizontal Digital's choices.

How a page is scored

A page is scored out of 100 points. Each check has a tier: essential, recommended or bonus. Essential checks share 80 of the points, recommended checks share 20, and bonus checks can add up to 5, with the total capped at 100. A check that does not apply to a page is skipped, and its points are shared among the rest.

One check is a gate and carries no points: a page marked noindex, nosnippet or max-snippet:0 is treated as blocked. Those three values tell search engines not to list a page or not to quote from it. The score then gets one of three labels.

In the demonstration Khushboo Sorthiya opened an article on a demo site called BrandX. The app gave the page 69, labeled almost there, with its checks listed under the four questions. After a meta description (the page's short summary for search engines) that was too short was rewritten, the score went up. A site has a score of its own from the 12 site checks: BrandX's was 85, labeled agent-ready. She did not say what the site's score of 85 is out of.

One check, as the session pictured it

"Content is there without JavaScript" is the first check under Read and an essential one. A slide in the session, headed "What the bot actually sees", shows the case it looks for: a page that a script assembles in the visitor's browser. The visitor sees the finished page. GPTBot, OpenAI's crawler, receives a body that holds an empty div and a script tag.

<html>
  <head><title>BrandX</title></head>
  <body>
    <div id="root"></div>
    <script src="/static/main.js"></script>
  </body>
</html>

The app compares the words in a page's draft with the words in the HTML the server returns. If at least 95% of the draft's words are there the check passes, and 80% earns half the check's points.

The 20 page checks

Each check sits under one of the four questions. The middle column restates the pass mark in the app's documentation. Where a check can be partly passed, the share it earns is a share of that check's own points.

The table uses several terms. A canonical link is the link in which a page names its own address. Markdown is a plain-text format, requested with Accept: text/markdown, and page chrome is the menus, headers and footers around a page's content. PageSpeed is Google's test of how fast a page loads, with a score out of 100. JSON-LD is a block of data in a page's HTML that states what the page describes, and a breadcrumb list is the trail of links that shows where a page sits in its site. Open Graph tags are a page's sharing tags. A reading grade is the U.S. school grade a reader needs to have reached to follow a text: 13 is the first year of college. Alt text is an image's written description.

The 12 site checks

These run once for the whole site and have their own score. Each has a tier, marked the same way. Two of them refer to seven search and user agents, which the second table names.

The terms here: robots.txt is the file that tells crawlers what they may fetch, and a sitemap lists a site's pages for them. Vary: Accept is a response header saying the answer depends on the format a request asks for. 404 and 410 are the status codes for a page that is not there or has been removed, and 200 is the code for a page that was found. MCP is the protocol an agent uses to call outside tools.

Open source

The session's last slide on the app gives the address of its repository and three steps: fork the app and install it as a custom app in SitecoreAI, add rules of your own, and send a pull request, or PR, for a check others would need.

https://github.com/horizontalintegration/is-it-agent-ready-for-sitecoreai

Every number, and who presented it

This section puts every figure in the report in one place, with what it measures and who presented it, so that a reader can check where a number came from. A last table lists the speakers and the sessions in which they spoke.

How to read it. The figures are grouped under the section that reports them, in the order they appear there. Where a section says the speaker or the slide did not give the base of a figure, the row says "base not stated". "On stage" means the figure was spoken and was not on a slide.

The new reader

Found is not chosen

How an assistant picks its sources

Before any data: four decisions

From finding to fix

Choosing the measure first

The agent-ready checklist

The pass marks of the checks, such as a title's length or a page's word count, are tabulated in the section on the checklist and are not repeated. The rows here are the figures of its scoring.

The speakers

These are the people the report credits, by company, with the session or sessions from which it reports them. All of the sessions were at Sitecore Symposium 2026 on Wednesday, November 11 or Thursday, November 12, 2026.